Variants in the CDKN2B and RTEL1 regions are associated with high-grade glioma susceptibility

eScholarship provides open access, scholarly publishingservices to the University of California and delivers a dynamicresearch platform to scholars worldwide.

Center for Bioinformatics and MolecularBiostatistics

UC San Francisco

Title:Variants in the CDKN2B and RTEL1 regions are associated with high-grade glioma susceptibility

Author:Xiao, Yuanyuan, University of California, San Francisco

Publication Date:07-08-2009

Publication Info:Center for Bioinformatics and Molecular Biostatistics, UC San Francisco

Permalink:http://escholarship.org/uc/item/32p785g8

Additional Info:in press

Keywords:adult glioma and genome-wide association study

Variants in the CDKN2B and RTEL1 regions areassociated with high-grade glioma susceptibilityMargaret Wrensch1,2,12, Robert B Jenkins3,12, Jeffrey S Chang4,12, Ru-Fang Yeh4,12, Yuanyuan Xiao4,Paul A Decker5, Karla V Ballman5, Mitchel Berger1, Jan C Buckner6, Susan Chang1, Caterina Giannini3,Chandralekha Halder3, Thomas M Kollmeyer3, Matthew L Kosel5, Daniel H LaChance7, Lucie McCoy1,Brian P O’Neill7, Joe Patoka1, Alexander R Pico8, Michael Prados1, Charles Quesenberry9, Terri Rice1,Amanda L Rynearson3, Ivan Smirnov1, Tarik Tihan10, Joe Wiemels2,4, Ping Yang11,13 & John K Wiencke1,2,13

The causes of glioblastoma and other gliomas remainobscure1,2. To discover new candidate genes influencingglioma susceptibility, we conducted a principal component–adjusted3 genome-wide association study (GWAS) of 275,895autosomal variants among 692 adult high-grade glioma cases(622 from the San Francisco Adult Glioma Study (AGS) and70 from the Cancer Genome Atlas (TCGA))4 and 3,992controls (602 from AGS and 3,390 from Illumina iControlDB(iControls)). For replication, we analyzed the 13 SNPs withP o 10�6 using independent data from 176 high-grade gliomacases and 174 controls from the Mayo Clinic. On 9p21,rs1412829 near CDKN2B had discovery P ¼ 3.4 � 10�8,replication P ¼ 0.0038 and combined P ¼ 1.85 � 10�10.On 20q13.3, rs6010620 intronic to RTEL1 had discoveryP ¼ 1.5 � 10�7, replication P ¼ 0.00035 and combinedP ¼ 3.40 � 10�9. For both SNPs, the direction of associationwas the same in discovery and replication phases.

Subject characteristics, including participation rates for the discoveryGWAS and replication phases, are listed in Supplementary Table 1a,b.The distribution of P values from the principal component–adjustedlogistic regression additive model across the genome for high-gradeglioma cases versus controls (Fig. 1) suggests potentially meaningfulassociations for several SNPs on chromosomes 1, 5, 9, 11 and 20.Supplementary Table 2 summarizes results for the 13 SNPs withP o 10�6 for association with high-grade glioma in discovery dataalong with results from replication data; SNPs with Hardy-WeinbergPo 10�5 in controls or with 45% missing data in any case or controlgroup were excluded. Three of these 13 SNPs (rs1412829 on 9p21, andrs6010620 and rs4809324 intronic to RTEL1 on 20q13.3) had

significant association with high-grade glioma risk in the discoveryphase (principal component analysis P o 1.8 � 10�7), were inde-pendent risk predictors in a multivariable analysis of 13 top hits, andwere replicated in the Mayo Clinic dataset (Table 1). As shown inTable 1 and Supplementary Table 2, the minor allele frequencies forthe three SNPs consistently differed in the same direction betweenhigh-grade glioma cases and controls regardless of data source (AGS,TCGA, iControls or Mayo Clinic). Supplementary Table 3 showsresults from the multivariable model of discovery data that includedall 13 SNPs (four from the 9p21 region, three in RTEL1, plus six othersin other locations). Eight SNPs, including one in the 9p21 region andtwo intronic to RTEL1, remained independently associated with high-grade glioma risk after adjustment for other SNPs. This was expectedgiven the strong linkage disequilibrium (LD) evident for the four 9p21SNPs and two of the three RTEL1 SNPs (Supplementary Table 4).

In discovery data, only the interaction between chromosome 9p21SNP rs1412829 and TERT SNP rs2736100 on chromosome 5 wasstatistically significant with Wald test P ¼ 0.019 (see Supplementary

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

1 × 10–12

1 × 10–10

1 × 10–8

1 × 10–6

1 × 10–4

1 × 10–2

1 3 4 5 6 7Chromosome

P v

alue

8 9 11 13 15 17 2021

Figure 1 Distribution of P values from principal component–adjusted logistic

regression additive model across the genome for high-grade glioma cases

versus controls. The 13 SNPs with P o 10�6 are shown in red.

Received 13 March; accepted 1 June; published online 5 July 2009; doi:10.1038/ng.408

1Department of Neurological Surgery, University of California, San Francisco, San Francisco, California, USA. 2Institute of Human Genetics, University of California,San Francisco, San Francisco, California, USA. 3Department of Experimental Pathology, Mayo Clinic, Rochester, Minnesota, USA. 4Department of Epidemiologyand Biostatistics, University of California, San Francisco, San Francisco, California, USA. 5Division of Biostatistics, 6Department of Oncology and 7Department ofNeurology, Mayo Clinic, Rochester, Minnesota, USA. 8Gladstone Institute of Cardiovascular Disease, University of California, San Francisco, San Francisco, California,USA. 9Division of Research, Kaiser Permanente, Oakland, California, USA. 10Department of Pathology, University of California, San Francisco, San Francisco,California, USA. 11Division of Epidemiology, Mayo Clinic, Rochester, Minnesota, USA. 12These authors contributed equally to this work. 13These authors jointly directedthe work. Correspondence should be addressed to M.W. ([email protected]).

NATURE GENETICS ADVANCE ONLINE PUBLICATION 1

LET TERS

Table 5 and its accompanying graph). However, close inspectionrevealed that the interaction probably resulted from an associationof the two SNPs in the controls.P values of SNPs on 9p21 and the LD plot (Fig. 2) show that the top

9p21 SNPs are located in or around CDKN2B. Haplotype analyses(Table 2) showed that a single haplotype for the four 9p21 SNPs wasmore common in cases than controls. Two haplotypes in RTEL1 wereassociated with increased and decreased risk, respectively (Table 2).The Mayo replication data also defined the identical haplotypeassociated with high-grade glioma risk for the four linked 9p21SNPs as identified during the discovery phase (Table 2). In addition,one of the two RTEL1 haplotypes identified in the discovery phase wasalso significantly associated in the replication samples (Table 2).Mantel-Haenszel combined P values for the UCSF and Mayo samplesfor the 13 SNPs are shown in Supplementary Table 6. The UCSFGWAS and Mayo replication suggest that regions of 9p21 (CDKN2B)and 20q13.3 (RTEL1) harbor SNPs associated with high-grade gliomarisk, as discussed further below.

The strongest and most consistent associations in the GWAS werewith a series of four SNPs within noncoding regions of the CDKN2Blocus on 9p21. CDKN2B lies adjacent to the well-known tumorsuppressor gene CDKN2A (encoding p16INK4A and p14ARF) in aregion that is frequently mutated, deleted or hypermethylated in awide variety of tumors, including high-grade glioma. The region iswithin 20 kb of constitutional deletions, including the hemizygousgermline deletion of CDKN2A that has been reported by the Mayogroup and others to be linked to familial melanoma and glioblastomasyndrome (Fig. 2)5. Mice with homozygous deletion of Cdkn2aand/or Cdkn2b are predisposed to develop tumors, including

gliomas6. CDKN2B, like CDKN2A, is a cyclin-dependent kinaseinhibitor which forms a complex with CDK4 or CDK6 and preventsthe activation of the cyclin-D–dependent kinases, thus regulating cellgrowth and cell cycle G1 progression. CDKN2B is frequently inacti-vated in glioma by homozygous deletion along with CDKN2A; 50–70% of primary high-grade gliomas show deletion of this region.Whereas tumor suppressor functions for CDKN2A have been firmlyestablished, only recently has CDKN2B been recognized as an effective‘backup’ for loss of CDKN2A7. In glioblastoma cells, overexpression ofCDKN2B in a CDKN2A-deficient background inhibited cell growth,induced replicative senescence and inhibited telomerase activity8. Incontrast to CDKN2A, CDKN2B is markedly induced by TGF-b. It hasthus been hypothesized that CDKN2B may be engaged under specialcircumstances, whereas CDKN2A plays a more general tumor sup-pressor function in response to DNA damage and hyperproliferativesignals8. TGF-b signaling information is relayed from the cell surfaceto the nucleus via the phosphorylation of SMAD proteins. A recentstudy identified a SMAD-binding region in the CDKN2B promoter; itis of interest that the SNPs associated with glioma in the current studyare in LD with the rs2069418 G4C SNP that lies in the crucialconserved 3¢ box adjacent to the SMAD binding element9. If a SNP inthis region reduces the responsiveness of CDKN2B to TGF-b, it couldallow cancer precursor cell populations to expand, thereby promotinggliomagenesis. It is unknown, however, whether any SNP in the regioncan affect TGF-b or any other cytokine signaling processes.

Although recent studies10–13 have identified chromosome 9p21 asan important region for coronary artery diseases (CAD) and type 2diabetes (T2D), the four glioma-associated chromosome 9p21SNPs are not in LD with SNPs associated with CAD or T2D

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Table 1 Three independent SNPs from the high-grade glioma discovery GWAS replicated in independent data from the Mayo Clinic

SNP rs1412829 rs6010620 rs4809324

Chromosome 9 20 20

Position 22033926 61780283 61788664

Gene symbol RTEL1 RTEL1

Minor allele C A C

Discovery set: UCSF AGS and TCGA cases and AGS and iControls

Number genotyped

Cases 692 692 692

Controls 3,989 3,991 3,979

Minor allele frequencies

Cases 0.47 0.17 0.15

Controls 0.39 0.23 0.10

Principal component–adjusted P value for cases versus controls 3.40 � 10–8 1.50 � 10–7 1.50 � 10–7

OR and 95% CI for 0, 1 or 2 minor alleles 1.39 (1.24–1.57) 0.68 (0.58–0.79) 1.54 (1.31–1.82)

Replication set: Mayo Clinic glioblastoma and anaplastic astrocytoma cases and controls

Number genotyped

Cases 175 175 176

Controls 173 174 174

Minor allele frequencies

Cases 0.53 0.15 0.16

Controls 0.41 0.26 0.10

Principal component–adjusted P value for cases versus controls 0.0038 0.00035 0.03

OR and 95% CI for 0, 1 or 2 minor alleles 1.56 (1.16–2.12) 0.48 (0.32–0.72) 1.66 (1.06–2.61)

Combined results

Mantel-Haenszel combined P values 1.85 � 10–10 3.40 � 10–9 1.70 � 10–9

OR and 95% CI 1.42 (1.27–1.58) 0.66 (0.57–0.76) 1.60 (1.37–1.87)

Principal component analysis implemented with EIGENSTRAT software. Complete results for 13 top hits with P o 10�6 from UCSF GWAS and Mayo Clinic replication P values areshown in Supplementary Table 2, with Mantel-Haenszel combined results presented in Supplementary Table 6.

2 ADVANCE ONLINE PUBLICATION NATURE GENETICS

LET TERS

(Supplementary Fig. 1). This suggests that separate regions onchromosome 9p21 may contribute to the risk of high-grade gliomaversus CAD and T2D.

Although replication results were not statistically significant, TERT isanother interesting gene identified in our GWAS; it encodes humantelomerase, which is a ribonucleoprotein polymerase that maintainstelomere ends by addition of the telomere repeat TTAGGG. TERTactivity is increased in glioblastoma14,15 and contributes to glioma cellgrowth16. A recent GWAS associated the region containing TERT with

idiopathic pulmonary fibrosis17. AnotherGWAS reported a significant associationbetween TERT SNP rs2736100 and lung can-cer18. On chromosome 20, we found that arelated gene, RTEL1, contains two SNPswithin intron 12 (rs6010620) and intron 17(rs4809324) significantly associated with high-grade glioma. RTEL1 is a DNA helicase critical

for regulation of telomere length in mice, and its loss is associated withshortened telomere length, chromosome breaks and translocations19.

This study is strengthened by the use of principal components3 toadjust for any residual population stratification after using severalquality control measures to assure that only unrelated subjects ofEuropean ancestry were included in the analyses (see Online Methodsfor details). Because glioma is a relatively rare disease, very largematched sets of glioma cases and controls are not currently availablefor GWAS. Consequently, we used a publicly available control group

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Table 2 Haplotype analysis of associations of high-grade glioma risk with SNPs in the 9p21 region and RTEL1

Cases (%) iControls (%) Odds ratioa P valuea

UCSF Adult Glioma Study (AGS) and the Cancer Genome Atlas high-grade glioma cases and AGS and Illumina controls

Chromosome 9p21: rs1063192, rs2157719, rs1412829, rs4977756

T-A-T-A 50.0 58.5 Referent

C-G-C-G 43.2 35.1 1.42 (1.26–1.60) 1.4 � 10�8

Rare haplotypesb 6.8 6.4 1.22 (0.96–1.56) 0.110

Global P valuea: 7.4 � 10�8

RTEL1: rs4809324, rs6010620, rs6089953

T-G-G 68.0 66.9 Referent

T-A-A 16.1 21.7 0.71 (0.61–0.83) 1.8 � 10�5

C-G-G 15.0 10.3 1.40 (1.18–1.66) 9.6 �10�5

Rare haplotypesb 0.9 1.1 1.12 (0.57–2.19) 0.750

Global P valuea: 3.6 � 10�9

Mayo Clinic glioblastoma and anaplastic astrocytoma cases and controls

Chromosome 9p21: rs1063192, rs2157719, rs1412829, rs4977756

T-A-T-A 44.6 56.6 Referent

C-G-C-G 49.7 36.5 1.68 (1.23–2.29) 0.001

Rare haplotypesb 5.7 6.9 0.96 (0.50–1.84) 0.900

Global P valuea: 0.002

RTEL1: rs4809324, rs6010620, rs6089953

T-G-G 68.1 63.2 Referent

T-A-A 14.8 25.0 0.52 (0.35–0.79) 0.002

C-G-G 16.5 10.3 1.42 (0.89–2.26) 0.144

Rare haplotypesb 0.6 1.5 0.41 (0.08–2.22) 0.303

Global P valuea: 0.002

aPrincipal component–adjusted odds ratios, confidence intervals and P values were estimated using EIGENSTRAT software; SNPs with individual P o 10�6 were included in the haplotype analyses.bRare haplotypes (o5%) were grouped together for these analyses.

chr9: 21800000

MTAP

CDKN2ACDKN2ACDKN2A

ANRILCDKN2B

BC038540

MTAP

7.5 5.0

3.3

1.7

0

5.0

2.5

UC

SF

–lo

g P

Mayo –log P

0

21900000 22000000

Family BFamily A

Other reported SNP disease associationsCoronary diseaseMyocardial infarction

Type 2 diabetes

Deletions in familial melanoma/glioblastoma syndrome (6)

22100000 22200000 22300000 22400000UCSC genes based on RefSeq, UniProt, GenBank, CCDS and comparative genomics

a

b

c

d

Figure 2 Map of the associated 9p21 region

in high-grade glioma. (a) Genes within region.

(b) Location of hemizygous deletion regions

previously linked to familial melanoma/

glioblastoma syndrome (blue)5. Also shown are

SNPs within the region that have been previously

reported to be associated with heart disease and

diabetes risk11. (c) –log P for SNPs within region;

note different scales for UCSF discovery phase

(blue bars, left x axis) and Mayo Clinic replication

phase (red bars, right x axis). P values are from

single point association tests of principal

component–adjusted additive logistic regression

of cases versus controls for 0, 1 or 2 minor

alleles. (d) LD of HapMap SNPs in region.

NATURE GENETICS ADVANCE ONLINE PUBLICATION 3

LET TERS

from Illumina to provide large numbers of controls for the discoveryphase. To minimize the possibility of false positives that might resultfrom using a nonmatched control set, we used carefully matched high-grade glioma cases and controls from the Mayo Clinic for replicationanalyses. In summary, this report shows that high-grade glioma risk isassociated with inherited variation in a region of 9p21 containingCDKN2B and a region of 20q13.3 tagged by two intronic SNPs inRTEL1. That the 9p21 region is frequently deleted in high-gradegliomas lends further biological plausibility to these findings.

METHODSMethods and any associated references are available in the onlineversion of the paper at http://www.nature.com/naturegenetics/.

Note: Supplementary information is available on the Nature Genetics website.

ACKNOWLEDGMENTSWork at University of California, San Francisco (UCSF) has been supported by USNational Institutes of Health grants R01CA52689 and UCSF Brain Tumor SPORE,P50CA097257, as well as by grants from the National Brain Tumor Foundation,the UCSF Lewis Chair in Brain Tumor Research and by donations from familiesand friends of J. Berardi, H. Glaser and E. Olsen. J.S.C. was also supported by afellowship from the National Cancer Institute (grant R25 CA 112355). Work at theMayo Clinic has been supported by the Mayo Clinic Brain Tumor SPORE (NIHP50 CA108961), the Mayo Clinic Comprehensive Cancer Center (P30 CA15083)and the Bernie and Edith Waterman Foundation. The San Francisco Adult GliomaStudy thanks the Northern California Cancer Center for identifying glioma cases;we also thank K. Aldape for pathology review and the pathology departmentsof Alexian, Alta Bates, Brookside, California Pacific, Doctors Pinole, Eden,El Camino, Good Samaritan, Highland, John Muir, Kaiser Redwood City,Kaiser San Francisco, Kaiser Santa Teresa, Los Gatos, Los Medanos, Marin General,Merrithew, Mills Peninsula, Mt. Diablo Hospital, Mt. Zion, Naval Hospital,O’Connor, Ralph K Davies, Saint Louise, San Francisco General, San Jose,San Leandro, San Mateo County, San Ramon Valley, Santa Clara Valley, Sequoia,Seton, St. Francis, St. Luke’s, St. Rose, Stanford, Summit, UC San Francisco, ValleyLivermore, Veterans Palo Alto, Veterans SF, and Washington Hospitals and MedicalCenters for providing tumor specimens for review. Genotyping services forSan Francisco study subjects were provided by deCODE Genetics, Iceland. Thecompany provided SNP and normalized CNV data and technical support indata analysis, including conference call tutorials in the use of the Disease MinerSoftware. We thank B. Scheithauer and C. Gianinni for their careful histologicalreview of all the primary high-grade gliomas collected at the Mayo Clinic for thisstudy. The Mayo Clinic Comprehensive Cancer Center Biospecimens andProcessing (TACMA), Gene Analysis, Biostatistics and Bioinformatics SharedResources were essential for the success of this study. We also thank K. Kelseyfor helpful suggestions on genotyping and interpretation of results, N. Risch forvery helpful suggestions on this paper and S. Sen for helpful discussions andsuggestions on statistical methods. Some computations were performed usingthe UCSF Biostatistics High Performance Computing System.

AUTHOR CONTRIBUTIONSM.W. was the overall UCSF study principal investigator who was responsible forsubject recruitment, oversaw all analyses and wrote parts of and synthesized thepaper. R.B.J. was the overall co-principal investigator of the Mayo study whooversaw the entire study (particularly laboratory quality control), interpreted theresults and wrote parts of the paper. J.S.C. was the UCSF epidemiologist whocontributed to development of the analysis plan, conducted statistical analyses andwrote parts of the paper. R.-F.Y. was the UCSF biostatistician who oversaw andconducted statistical analyses of the discovery phase and wrote parts of the paper.Y.X. was the UCSF biostatistician who conducted statistical analyses of thediscovery and combined phases and wrote parts of the paper. P.A.D. was the Mayostatistician who performed all Mayo data analysis. K.V.B. was the Mayo leadstatistician who participated in study design and the analysis plan. M.B.was the principal investigator of the UCSF Brain Tumor SPORE and a clinicalcollaborator who provided access for subject recruitment. J.C.B. was the Mayoneuro-oncologist who led subject recruitment. S.C. was the co-director of theUCSF neuro-oncology clinic who assisted in subject recruitment. C.G. was theMayo pathologist who verified all pathologic diagnosis of Mayo cases. C.H. wasthe Mayo laboratory technologist responsible for specimen preparation forgenotyping. T.M.K. was the Mayo laboratory manager responsible for specimen

storage and retrieval. M.L.K. provided statistical support for all Mayo analyses.D.H.L. was the Mayo neuro-oncologist who facilitated subject enrollment andmedical record data collection. L.M. was the UCSF project coordinator responsiblefor subject recruitment and preparation of datasets for analyses, and alsoconducted analyses. B.P.O. was the principal investigator of Mayo brain tumorSPORE and neurologist who facilitated subject enrollment and medical record datacollection. J.P. was the UCSF laboratory manager responsible for specimen storage,retrieval and preparation for genotyping. A.R.P. was the UCSF/Gladstonebioinformatician who participated in selecting the genotyping platform, developingthe analytical plan and reviewing the paper. M.P. was the co-director of the UCSFneuro-oncology clinic who assisted in subject recruitment. C.Q. participated insubject recruitment and pathology specimen accrual from Kaiser PermanenteNorthern California. T.R. was the UCSF project coordinator responsible for subjectrecruitment, prepared datasets for analyses, conducted analyses and wrote parts ofpaper. A.L.R. was the Mayo project coordinator responsible for subjectrecruitment. I.S. was the UCSF bioinformatician who participated in developingthe analytical plan, data analysis and interpreting results. T.T. was the UCSFneuropathologist who participated in subject identification, accrual anddevelopment of the analytical strategy. J.W. was the UCSF epidemiologist whoparticipated in choice of genotyping platform and development of the analyticalstrategy and oversaw sample preparation. P.Y. was the overall co-principalinvestigator of the Mayo study who oversaw the entire study (particularly studydesign for subject recruitment, control enrollment, data quality control andanalyses), interpreted results and wrote parts of the paper. J.K.W. was the UCSFstudy co-principal investigator who oversaw all aspects of laboratory work,participated in study design, subject accrual and development of the analysis plan,and wrote the discussion portion of the paper.

Published online at http://www.nature.com/naturegenetics/.

Reprints and permissions information is available online at http://npg.nature.com/

reprintsandpermissions/.

1. CBTRUS. Primary Brain Tumors in the United States, Statistical Report 2000–2004 (Central Brain Tumor Registry of the United States, Chicago, Illinois, 2008).

2. Schwartzbaum, J.A., Fisher, J.L., Aldape, K.D. & Wrensch, M. Epidemiology andmolecular pathology of glioma. Nat. Clin. Pract. Neurol. 2, 494–503 (2006).

3. Price, A.L. et al. Principal components analysis corrects for stratification in genome-wide association studies. Nat. Genet. 38, 904–909 (2006).

4. Cancer Genome Atlas Research Network. Comprehensive genomic characterizationdefines human glioblastoma genes and core pathways. Nature 455, 1061–1068(2008).

5. Bahuau, M. et al. Germ-line deletion involving the INK4 locus in familial proneness tomelanoma and nervous system tumors. Cancer Res. 58, 2298–2303 (1998).

6. Sharpless, N.E. et al. Loss of p16Ink4a with retention of p19Arf predisposes mice totumorigenesis. Nature 413, 86–91 (2001).

7. Krimpenfort, P. et al. p15Ink4b is a critical tumour suppressor in the absence ofp16Ink4a. Nature 448, 943–946 (2007).

8. Fuxe, J. et al. Adenovirus-mediated overexpression of p15INK4B inhibits humanglioma cell growth, induces replicative senescence, and inhibits telomerase activitysimilarly to p16INK4A. Cell Growth Differ. 11, 373–384 (2000).

9. Seoane, J. et al. TGFb influences Myc, Miz-1 and Smad to control the CDK inhibitorp15INK4b. Nat. Cell Biol. 3, 400–408 (2001).

10. Lemmens, R. et al. Variant on 9p21 strongly associates with coronary heart disease,but lacks association with common stroke. Eur. J. Hum. Genet. advance onlinepublication, doi:10.1038/ejhg.2009.42 (25 March 2009).

11. Mohlke, K.L., Boehnke, M. & Abecasis, G.R. Metabolic and cardiovascular traits: anabundance of recently identified common genetic variants. Hum. Mol. Genet. 17,R102–R108 (2008).

12. Schaefer, A.S. et al. Identification of a shared genetic susceptibility locus for coronaryheart disease and periodontitis. PLoS Genet. 5, e1000378 (2009).

13. Schunkert, H. et al. Repeated replication and a prospective meta-analysis of theassociation between chromosome 9p21.3 and coronary artery disease. Circulation117, 1675–1684 (2008).

14. Shervington, A. et al. Glioma: what is the role of c-Myc, hsp90 and telomerase? Mol.Cell. Biochem. 283, 1–9 (2006).

15. Maes, L. et al. Relation between telomerase activity, hTERT and telomere length forintracranial tumours. Oncol. Rep. 18, 1571–1576 (2007).

16. Falchetti, M.L. et al. Telomerase inhibition impairs tumor growth in glioblastomaxenografts. Neurol. Res. 28, 532–537 (2006).

17. Mushiroda, T. et al. A genome-wide association study identifies an association of acommon variant in TERT with susceptibility to idiopathic pulmonary fibrosis. J. Med.Genet. 45, 654–656 (2008).

18. McKay, J.D. et al. Lung cancer susceptibility locus at 5p15.33. Nat. Genet. 40,1404–1406 (2008).

19. Ding, H. et al. Regulation of murine telomere length by Rtel: an essential geneencoding a helicase-like protein. Cell 117, 873–886 (2004).

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

4 ADVANCE ONLINE PUBLICATION NATURE GENETICS

LET TERS

ONLINE METHODSEthics. These studies were approved by the University of California

San Francisco Committee on Human Research and Mayo Clinic Office for

Human Research Protection. Informed consent was obtained from all

study participants.

Study subjects for discovery phase. Genotyping data came from four groups

of subjects: 622 high-grade astrocytic glioma cases and 602 controls from AGS,

3,390 controls from Illumina controls (iControls) and 70 glioblastoma cases

from TCGA4 (Supplementary Table 1a) that passed quality control measures

described below, including checks for relatedness and European ancestry.

Details of subject recruitment for AGS have been provided previously20,21.

Briefly, cases aged 20 or older diagnosed with histologically confirmed incident

gliomas (International Classification of Diseases for Oncology, morphology

codes 9380-9481) were recruited from the local population-based registry, the

Northern California Rapid Case Ascertainment program and the University of

California, San Francisco Neuro-oncology clinic between 1997 and 2006.

Additional pathology reviews were conducted by specialty trained neuropathol-

ogists. Glioblastoma, which is the diagnosis for the large majority (84%) of

cases, is a diagnosis with good concordance between pathologists22. Although

survival bias is a concern for studies of glioblastoma, we obtained blood from

subjects within a median of 80 days from diagnosis. Nevertheless, the results

may not apply to those with the most rapidly fatal forms of this disease. AGS

controls aged 20 years or older from the same residential area as cases were

identified using random digit dialing and were frequency matched to cases on

age, sex and ancestry. Consenting participants provided blood and/or buccal

specimens and information during an in-person or telephone interview.

Because of the large-scale genotyping platform used, only subjects who

provided blood specimens were included in the present analysis. We initially

only included individuals who self-identified as white in the genotyping, but

then used methods described below to verify European ancestry.

We also assembled an independent control genotype dataset of 3,390

nonredundant European-ancestry controls from Illumina iControlDB. The

subjects are anonymous, with information only on their age, sex and ancestry.

The iControl data also included 262 HapMap samples (30 CEU parent-child

trios (Utah residents with ancestry from northern and western Europe), 84 YRI

(Yoruba in Ibadan, Nigeria) and 88 Chinese or Japanese) that we used to

identify and remove subjects with evidence of non-European ancestry from our

analysis. We checked for evidence of non-European ancestry (Supplementary

Fig. 2) and sample duplicates or related subjects (IBS 4 1.6; Supplementary

Fig. 3) among AGS samples, TCGA and iControls by performing multi-

dimensional scaling (MDS) analysis on 20 bootstrap samples of 1,000 random

autosomal biallelic SNPs. Following these quality assessment measures, we

obtained a total of 3,390 European-ancestry controls from three different

Illumina panels with up to 306,154 autosomal SNPs overlapping the Human-

Hap370duo panel used for the AGS subjects: Illumina HumanHap300 (n¼ 336

subjects), HumanHap550v1 (n ¼ 1,519) and HumanHap550v3 (n ¼ 1,552).

We downloaded HumanHap550 platform genotyping data from blood

specimen DNA and demographic data for 89 glioblastoma cases from the

Cancer Genome Atlas (TCGA)4. Although 72 were identified as white, our

analyses showed that one had non-European ancestry (Supplementary Fig. 3)

and another appeared to duplicate an AGS case, leaving 70 TCGA cases.

Sample preparation and genotyping for discovery phase and quality control.

DNA was isolated from whole blood using Gentra Puregene DNA isolation kit

(Qiagen) and quantified using Picogreen reagent (Invitrogen). Genotyping was

conducted by deCODE Genetics. Samples were randomized before plating on

specimen plates provided by deCODE Genetics. The genotyping assay panel

used was Illumina’s HumanCNV370-Duo BeadChip. For this paper, we only

analyzed autosomal SNPs. A complete list of the SNPs on this panel is available

either from Illumina or on the publicly available website of SNPLogic23. In

addition to randomization of samples and the quality control measures

provided by deCODE Genetics, we included two duplicate samples per plate

and one CEPH24 trio (parents and child) per plate. DNA was re-extracted for

any samples with call rates o98% and genotyped again and only samples

reaching call rate Z98% were included in these analyses. We genotyped a total

of 1,403 samples including AGS high-grade glioma cases (n ¼ 623) and

controls (n ¼ 602), duplicates (n ¼ 51), some subjects that were determined

to be ineligible based on self-described ancestry (n ¼ 22), AGS cases with non-

high-grade glioma histologies (n ¼ 67) and CEPH samples we provided for

quality control (n¼ 36); one sample was deleted because of inadequate call rate

and one because of a mismatch between stated and genotyped sex. In addition,

one case subject was removed because of not clustering with those of European

ancestry in the identity-by-state (IBS) analysis (Supplementary Fig. 2). Thus,

we used genotyping data from 622 AGS high-grade glioma cases and 602 AGS

controls (Supplementary Table 1a).

The assay panel contained a total of 370,404 probes, of which 353,202 were

associated with a reference sequence number and 17,202 with a copy number

variant. There were a total of 342,554 SNPs provided in the genotyping files

received from deCODE, with 331,697 autosomal SNPs. Of these, 250 had

completely missing data and 4,941 were nonpolymorphic; this left 326,506

biallelic SNPs for analysis.

Quality control information for AGS cases and controls is presented in

Supplementary Figures 4, 5 and 6 showing autosomal heterozygosity for all

1,403 genotyped samples, percent of SNPs missing in cases versus controls, and

Hardy-Weinberg equilibrium P values for AGS cases versus AGS controls. We

used this information to exclude SNPs with poorer data quality from presented

data. Similar quality measures were computed and used for filtering iControl

and TCGA data (data not shown). After excluding SNPs with P o 10�5 for

Hardy-Weinberg equilibrium in either AGS controls or iControls and those

with 45% missing genotyping data in any of the four subject groups, AGS

cases or controls, iControls or TCGA cases, there were 275,895 SNPs to

consider in case-control association tests.

In addition to ancestry checks described above, we ran EIGENSTRAT3 to

adjust for other possible population or batch differences between the combined

high-grade glioma cases and control groups. The quantile-quantile plot

(Supplementary Fig. 7) for the EIGENSTRAT-adjusted P values comparing

AGS and TCGA cases versus AGS and Illumina controls showed good

correspondence between observed and expected test P values for equality of

allele frequencies for the vast majority of the SNPs except for the ones with very

low P values. The genomic control parameter, 1.058, was very similar to that

found by Hung et al.25 of 1.03. They interpreted this to indicate there was no

systematic increase in false-positive findings owing to population stratification

or any other form of bias. We also present the quantile-quantile plot comparing

principal component–adjusted P values for AGS versus Illumina controls

(Supplementary Fig. 8). The genomic control parameter for this comparison

is 1.07.

Statistical methods for discovery phase. We used three software packages to

conduct all analyses: R, Disease Miner (deCODE genetics) and EIGENSTRAT3.

We also used Microsoft Excel and SAS for additional data manipulations and

visualization. As noted above, quality control analyses included computation

of sample heterozygosity, percent missing data (no genotype call) and

Hardy-Weinberg equilibrium. The primary analyses used in this paper were

EIGENSTRAT-adjusted single point association results from the additive

logistic regression model for 0, 1 or 2 copies of the minor allele (equivalent

to a Cochran-Armitage test for trend). We used a two-stage statistical approach

to identify SNPs independently associated with high-grade glioma and poten-

tial SNP-SNP interaction effects. In the first stage, a backward selection

procedure was used to obtain the best logistic regression model, using the 13

SNPs that produced a P value o 10�6 in the single locus models. Eight SNPs

remained in the model after the backward selection procedure, suggesting a

significant and independent association with high-grade glioma risk, among

which one 9p21 SNP (out of four) and two RTEL1 SNPs (out of three) were

retained (Supplementary Table 3). Having confirmed significant main effects

for these eight SNPs, we investigated in the second stage whether there are SNP-

SNP pairwise interactions among them. This was performed again in a

backward selection logistic regression framework using the eight SNPs and

all pairwise interactions among them as covariates. Only the interaction

between the 9p21 SNP rs1412829 and TERT SNP rs2736100 was statistically

significant with P o 0.05 (Wald test) after backward selection. In Supplemen-

tary Table 5, we calculated ORs of the nine genotype groups of rs1412829

[TT,CT,CC] and rs2736100 [GG,GT,TT] using [rs1412829 ¼ TT, rs2736100 ¼GG] as a referent group. Because the resulting P values are subject to multiple

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

NATURE GENETICS doi:10.1038/ng.408

testing errors, we also provide Bonferroni-adjusted P values. For the top hits,

we also conducted sensitivity analyses using logistic additive regression models

adjusting for age and sex (data not shown).

Haplotype analyses for the chromosome 9 and RTEL1 SNPs were carried out

using the haplo.stats R package. Rare haplotypes (o5%) were combined into

one group for analysis. Global P values were calculated to assess whether any

one of the haplotypes was over- or underrepresented among cases compared to

controls. Haplotype trend regression was conducted to calculate OR associated

with each copy of a specific haplotype using the most frequent haplotype as the

referent group (Table 2). LD (r2) among the four significant 9p21 SNPs and

three significant RTEL1 SNPs was calculated using PROC ALLELE procedure in

SAS Genetics. A LD plot (Supplementary Fig. 1) showing the linkage pattern

in r2 between our four 9p21 glioma-associated SNPs and seven CAD or T2D

SNPs10–13 was constructed using Haploview26 with HapMap data of

European ancestry (CEU). As shown, none of the four glioma-associated SNPs

(rs1063192, rs2157719, rs1412829, rs4977756 at positions 21993367 to

22058652) shows any marked LD with seven CAD- or T2D-associated SNPs

(rs2891168, rs1333042, rs2383206, rs10757278, rs1333048, rs1333049,

rs2383208 at positions 22088618 to 22122075), with maximum r2 ¼ 0.47

between rs4977756 and rs2891168. For rs1412829, the 9p21 SNP independently

associated with glioma risk in multivariate analysis, the maximum correlation

with the CAD- or T2D-associated SNP (rs2891168) was r2 ¼ 0.23.

For completeness for other glioma genetics researchers, we also present

SNP-disease association data for any SNP with P o 0.001 for tests of

associations with high-grade glioma (Supplementary Table 7). We also

conducted all analyses on glioblastoma only cases; no major differences for

statistically significant SNPs were found (data not shown).

To obtain combined estimates of high-grade glioma risk for top SNPs from

UCSF and Mayo, we used the Mantel-Haenszel method to estimate the OR,

95% CI and P value; the test is for equality of allele frequencies between cases

and controls.

Subjects for replication phase. The Mayo Clinic case group included 176

individuals with glioblastoma and anaplastic astrocytomas newly diagnosed

between 2005 and 2008. Cases were identified within 24 h of diagnosis, except

for those who had their initial diagnosis elsewhere, followed by verification at

the Mayo Clinic. The cases consisted of 67 (38%) women and 109 (62%) men

who were 53.8 ± 12.6 years old; 174 (98%) were white; 114 (65%) had

glioblastoma; 62 (35%) had anaplastic astrocytomas. Pathologic diagnosis was

confirmed by review of the primary surgical material for all cases by two Mayo

Clinic neuropathologists based on surgically resected material. The control

group consisted of consented individuals who had a general medical exam at

the Mayo Clinic. Matching variables were sex, date of birth (within two and one

half years), self-identified race (Hispanic white, non-Hispanic white, American

Indian, African American, Asian, Pacific Islander, Other) and residence.

Geographic region of residence was matched in three zones based on the

distance to the Mayo Clinic Rochester: Olmsted County; the rest of Minnesota,

Wisconsin, Iowa, North Dakota and South Dakota; and the rest of the United

States and Canada. Excluded were individuals under the age of 18 and those

with a history of brain tumor. The Mayo Clinic case and control enrollment

research protocol was approved by Mayo Institutional Review Board. These

cases and controls were genotyped using Illumina 610Quad arrays.

Sample preparation and genotyping for replication phase and quality

control. DNA was isolated from snap-frozen, buffy-coat samples using an

AutoGenFlex STAR system (AutoGen) with Qiagen’s FlexiGene DNA AGF3000

kit and AutoGen’s blood DNA finishing kit. DNA was quantified using a

ND-1000 spectrophotometer (Thermo Scientific) and normalized to 50 ng/ml

using 10 mM Tris HCl, 0.1 mM EDTA, pH 8.0 buffer (Teknova). Genotyping

was performed using Illumina 610Quad SNP arrays (Illumina) according to the

manufacturer’s recommendations. Briefly, 200 ng of genomic DNA was

amplified then fragmented. The fragmented DNA was hybridized on Illumina’s

Human 610-Quad BeadChip. Fluorescent labeling was performed by single-

base extension using labeled nucleotides. The BeadChip was then scanned with

Illumina’s Bead Array Reader. Samples, including positive controls, were

processed in a 96-well format.

We carried out allele calling using Illumina’s Genotyping Module version

3.3.7 in BeadStudio version 3.1.3.0. We summarized concordance in interplate,

intraplate and overall subject replicates to investigate potential genotyping

error. Subject-level call rates were calculated and those subjects with call rates

o0.9 were excluded from further analysis. Individual SNP call rates were

summarized and SNPs with call rates o0.9 were excluded from the analysis.

The minor allele frequency (MAF) was calculated for each SNP, and SNPs with

MAF o 0.01 were excluded from further analysis. The above analyses were

done on the complete set of data, and each analysis was repeated separately for

each plate to investigate any potential plate effects. The overall Illumina subject

call rate across all SNPs for Mayo Clinic cases and controls was 97.5 ± 0.02

(median 98.3; range 90.0–98.4). Inter- and intraplate replicate analysis was

performed for the 13 SNPs summarized in Table 1. For all 13 SNPs, all inter-

and intraplate replicates were identical.

Statistical methods for replication phase. The frequency distribution at each

SNP locus was tested against the Hardy-Weinberg equilibrium (HWE) under

the allele mendelian biallelic expectation using the w2 test. SNPs with HWE

P values o0.001 for control subjects were excluded from the analysis. The

principal component approach was implemented in EIGENSTRAT to deter-

mine whether there was any evidence of population stratification in the Mayo

cases and controls3,27. We used an additive logistic regression model for 0, 1 or

2 copies of the minor allele for candidate SNPs to investigate the association of

glioma risk. Significant principal components from the population stratifica-

tion analysis were included as covariates in the logistic regression models.

Haplotype blocks were estimated using Haploview26. The multiple SNP

marker-disease association by estimated haplotype was evaluated using

haplo.score (a software developed by the Mayo Clinic), which accounts for

ambiguous linkage phase28.

20. Felini, M.J. et al. Reproductive factors and hormone use and risk of adult gliomas.Cancer Causes Control 20, 87–96 (2009).

21. Wrensch, M. et al. Nonsynonymous coding single-nucleotide polymorphisms spanningthe genome in relation to glioblastoma survival and age at diagnosis. Clin. Cancer Res.13, 197–205 (2007).

22. Davis, F.G. et al. Issues of diagnostic review in brain tumor studies: from the BrainTumor Epidemiology Consortium. Cancer Epidemiol. Biomarkers Prev. 17, 484–489(2008).

23. Pico, A.R. et al. SNPLogic: an interactive single nucleotide polymorphism selection,annotation, and prioritization system. Nucleic Acids Res. 37, D803–D809 (2009).

24. Dausset, J. et al. Centre d’etude du polymorphisme humain (CEPH): collaborativegenetic mapping of the human genome. Genomics 6, 575–577 (1990).

25. Hung, R.J. et al. A susceptibility locus for lung cancer maps to nicotinic acetylcholinereceptor subunit genes on 15q25. Nature 452, 633–637 (2008).

26. Barrett, J.C., Fry, B., Maller, J. & Daly, M.J. Haploview: analysis and visualization of LDand haplotype maps. Bioinformatics 21, 263–265 (2005).

27. Patterson, N., Price, A.L. & Reich, D. Population structure and eigenanalysis. PLoSGenet. 2, e190 (2006).

28. Schaid, D.J., Rowland, C.M., Tines, D.E., Jacobson, R.M. & Poland, G.A. Score testsfor association between traits and haplotypes when linkage phase is ambiguous. Am. J.Hum. Genet. 70, 425–434 (2002).

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

doi:10.1038/ng.408 NATURE GENETICS